Combining feature fusion and attention mechanism for face image restoration

Abstract

We propose a face image restoration method that combines feature fusion and attention mechanisms for the current image restoration field that generates blurred images, artifacts, inconsistent texture and structure fusion. The model divides image restoration into two stages. First, the edge information repaired by the edge generation adversarial network is used as the prior knowledge of the image, and then the generated prior knowledge and the broken image are put into the image repair network to generate the complete image. We introduce a texture-structure feature fusion method in the generator structure to solve the texture and structure fusion inconsistency problem and use a dense residual layerhopping connection to mitigate the gradient disappearance problem while speeding up the model convergence and introduce a spatial and channel attention mechanism to generate correct semantic connections to enhance the model performance and suppress image blurring. We apply the algorithm to the CelebA-HQ face dataset, and compared with the current mainstream restoration algorithms, quantitative analysis shows that the method in this paper outperforms in three metrics, PSNR, SSIM, and L1.